feat: Compatible with XPU monitoring (#11088)

ssongliu · web-flow · commit 8e03b24b0c76 · 2025-11-26T22:30:15.000+08:00
diff --git a/agent/app/dto/monitor.go b/agent/app/dto/monitor.go
@@ -44,6 +44,7 @@ type MonitorGPUSearch struct {
 	EndTime     time.Time `json:"endTime"`
 }
 type MonitorGPUData struct {
+	GPUType          string                 `json:"gpuType"`
 	ProductNames     []string               `json:"productNames"`
 	Date             []time.Time            `json:"date"`
 	GPUValue         []float64              `json:"gpuValue"`
@@ -58,8 +59,8 @@ type GPUPowerUsageHelper struct {
 	Percent float64 `json:"percent"`
 }
 type GPUMemoryUsageHelper struct {
-	Total   int     `json:"total"`
-	Used    int     `json:"used"`
+	Total   float64 `json:"total"`
+	Used    float64 `json:"used"`
 	Percent float64 `json:"percent"`
 
 	GPUProcesses []GPUProcess `json:"gpuProcesses"`
diff --git a/agent/app/model/monitor.go b/agent/app/model/monitor.go
@@ -39,8 +39,8 @@ type MonitorGPU struct {
 	Temperature   int     `json:"temperature"`
 	PowerDraw     float64 `json:"powerDraw"`
 	MaxPowerLimit float64 `json:"maxPowerLimit"`
-	MemUsed       int     `json:"memUsed"`
-	MemTotal      int     `json:"memTotal"`
+	MemUsed       float64 `json:"memUsed"`
+	MemTotal      float64 `json:"memTotal"`
 	FanSpeed      int     `json:"fanSpeed"`
 	Processes     string  `json:"processes"`
 }
diff --git a/agent/app/service/monitor.go b/agent/app/service/monitor.go
@@ -131,17 +131,21 @@ func (m *MonitorService) LoadGPUMonitorData(req dto.MonitorGPUSearch) (dto.Monit
 	}
 	if len(req.ProductName) == 0 {
 		if gpuExist {
+			data.GPUType = "gpu"
 			gpuInfo, err := gpuclient.LoadGpuInfo()
 			if err != nil || len(gpuInfo.GPUs) == 0 {
+				global.LOG.Error("Load GPU info failed or no GPU found, err: ", err)
 				return data, buserr.New("ErrRecordNotFound")
 			}
 			req.ProductName = gpuInfo.GPUs[0].ProductName
 			for _, item := range gpuInfo.GPUs {
 				data.ProductNames = append(data.ProductNames, item.ProductName)
 			}
 		} else {
+			data.GPUType = "xpu"
 			xpuInfo, err := xpuClient.LoadGpuInfo()
 			if err != nil || len(xpuInfo.Xpu) == 0 {
+				global.LOG.Error("Load XPU info failed or no XPU found, err: ", err)
 				return data, buserr.New("ErrRecordNotFound")
 			}
 			req.ProductName = xpuInfo.Xpu[0].Basic.DeviceName
@@ -159,15 +163,18 @@ func (m *MonitorService) LoadGPUMonitorData(req dto.MonitorGPUSearch) (dto.Monit
 		data.Date = append(data.Date, gpu.CreatedAt)
 		data.GPUValue = append(data.GPUValue, gpu.GPUUtil)
 		data.TemperatureValue = append(data.TemperatureValue, gpu.Temperature)
-		data.PowerValue = append(data.PowerValue, dto.GPUPowerUsageHelper{
-			Total:   gpu.MaxPowerLimit,
-			Used:    gpu.PowerDraw,
-			Percent: gpu.PowerDraw / gpu.MaxPowerLimit * 100,
-		})
+		powerItem := dto.GPUPowerUsageHelper{
+			Total: gpu.MaxPowerLimit,
+			Used:  gpu.PowerDraw,
+		}
+		if powerItem.Total != 0 {
+			powerItem.Percent = powerItem.Used / powerItem.Total
+		}
+		data.PowerValue = append(data.PowerValue, powerItem)
 		memItem := dto.GPUMemoryUsageHelper{
 			Total:   gpu.MemTotal,
 			Used:    gpu.MemUsed,
-			Percent: float64(gpu.MemUsed) / float64(gpu.MemTotal) * 100,
+			Percent: gpu.MemUsed / gpu.MemTotal * 100,
 		}
 		var process []dto.GPUProcess
 		if err := json.Unmarshal([]byte(gpu.Processes), &process); err == nil {
@@ -564,14 +571,13 @@ func saveGPUDataToDB() {
 	var list []model.MonitorGPU
 	for _, gpuItem := range gpuInfo.GPUs {
 		item := model.MonitorGPU{
-			ProductName:   gpuItem.ProductName,
-			GPUUtil:       loadGPUInfoFloat(gpuItem.GPUUtil),
-			Temperature:   loadGPUInfoInt(gpuItem.Temperature),
-			PowerDraw:     loadGPUInfoFloat(gpuItem.PowerDraw),
-			MaxPowerLimit: loadGPUInfoFloat(gpuItem.MaxPowerLimit),
-			MemUsed:       loadGPUInfoInt(gpuItem.MemUsed),
-			MemTotal:      loadGPUInfoInt(gpuItem.MemTotal),
-			FanSpeed:      loadGPUInfoInt(gpuItem.FanSpeed),
+			ProductName: gpuItem.ProductName,
+			GPUUtil:     loadGPUInfoFloat(gpuItem.GPUUtil),
+			Temperature: loadGPUInfoInt(gpuItem.Temperature),
+			PowerDraw:   loadGPUInfoFloat(gpuItem.PowerDraw),
+			MemUsed:     loadGPUInfoFloat(gpuItem.MemUsed),
+			MemTotal:    loadGPUInfoFloat(gpuItem.MemTotal),
+			FanSpeed:    loadGPUInfoInt(gpuItem.FanSpeed),
 		}
 		process, _ := json.Marshal(gpuItem.Processes)
 		if len(process) != 0 {
@@ -596,25 +602,28 @@ func saveXPUDataToDB() {
 	var list []model.MonitorGPU
 	for _, xpuItem := range xpuInfo.Xpu {
 		item := model.MonitorGPU{
-			ProductName: xpuItem.Basic.DeviceName,
-			GPUUtil:     loadGPUInfoFloat(xpuItem.Stats.MemoryUtil),
-			Temperature: loadGPUInfoInt(xpuItem.Stats.Temperature),
-			PowerDraw:   loadGPUInfoFloat(xpuItem.Stats.Power),
-			MemUsed:     loadGPUInfoInt(xpuItem.Stats.MemoryUsed),
-			MemTotal:    loadGPUInfoInt(xpuItem.Basic.Memory),
-		}
-		var processItem []dto.GPUProcess
-		for _, ps := range xpuItem.Processes {
-			processItem = append(processItem, dto.GPUProcess{
-				Pid:         fmt.Sprintf("%v", ps.PID),
-				Type:        ps.SHR,
-				ProcessName: ps.Command,
-				UsedMemory:  ps.Memory,
-			})
-		}
-		process, _ := json.Marshal(processItem)
-		if len(process) != 0 {
-			item.Processes = string(process)
+			ProductName:   xpuItem.Basic.DeviceName,
+			GPUUtil:       loadGPUInfoFloat(xpuItem.Stats.MemoryUtil),
+			Temperature:   loadGPUInfoInt(xpuItem.Stats.Temperature),
+			PowerDraw:     loadGPUInfoFloat(xpuItem.Stats.Power),
+			MaxPowerLimit: float64(xpuItem.Config.PowerLimit),
+			MemUsed:       loadGPUInfoFloat(xpuItem.Stats.MemoryUsed),
+			MemTotal:      loadGPUInfoFloat(xpuItem.Basic.Memory),
+		}
+		if len(xpuItem.Processes) != 0 {
+			var processItem []dto.GPUProcess
+			for _, ps := range xpuItem.Processes {
+				processItem = append(processItem, dto.GPUProcess{
+					Pid:         fmt.Sprintf("%v", ps.PID),
+					Type:        ps.SHR,
+					ProcessName: ps.Command,
+					UsedMemory:  ps.Memory,
+				})
+			}
+			process, _ := json.Marshal(processItem)
+			if len(process) != 0 {
+				item.Processes = string(process)
+			}
 		}
 		list = append(list, item)
 	}
@@ -633,6 +642,7 @@ func loadGPUInfoInt(val string) int {
 }
 func loadGPUInfoFloat(val string) float64 {
 	valItem := strings.ReplaceAll(val, "W", "")
+	valItem = strings.ReplaceAll(valItem, "MB", "")
 	valItem = strings.ReplaceAll(valItem, "%", "")
 	valItem = strings.TrimSpace(valItem)
 	data, _ := strconv.ParseFloat(valItem, 64)
diff --git a/agent/init/migration/migrations/init.go b/agent/init/migration/migrations/init.go
@@ -721,7 +721,7 @@ var UpdateDatabase = &gormigrate.Migration{
 }
 
 var AddGPUMonitor = &gormigrate.Migration{
-	ID: "20251119-add-gpu-monitor",
+	ID: "20251127-add-gpu-monitor",
 	Migrate: func(tx *gorm.DB) error {
 		return global.GPUMonitorDB.AutoMigrate(&model.MonitorGPU{})
 	},
diff --git a/agent/utils/ai_tools/xpu/xpu_info.go b/agent/utils/ai_tools/xpu/xpu_info.go
@@ -10,6 +10,7 @@ type XpuInfo struct {
 type Xpu struct {
 	Basic     Basic     `json:"basic"`
 	Stats     Stats     `json:"stats"`
+	Config    Config    `json:"config"`
 	Processes []Process `json:"processes"`
 }
 
@@ -23,6 +24,11 @@ type Basic struct {
 	PciBdfAddress string `json:"pciBdfAddress"`
 }
 
+type Config struct {
+	PowerLimit      int    `json:"power_limit"`
+	PowerValidRange string `json:"power_vaild_range"`
+}
+
 type Stats struct {
 	Power       string `json:"power"`
 	Frequency   string `json:"frequency"`
diff --git a/frontend/src/api/interface/host.ts b/frontend/src/api/interface/host.ts
@@ -167,6 +167,7 @@ export namespace Host {
         endTime: Date;
     }
     export interface MonitorGPUData {
+        gpuType: string;
         productNames: Array<string>;
         date: Array<Date>;
         gpuValue: Array<number>;
diff --git a/frontend/src/lang/modules/zh.ts b/frontend/src/lang/modules/zh.ts
@@ -656,7 +656,7 @@ const message = {
         gpu: {
             gpu: 'GPU 监控',
             base: '基础信息',
-            gpuHelper: '当前系统未检测到 NVIDIA-SMI或者XPU-SMI 指令，请检查后重试！',
+            gpuHelper: '当前系统未检测到 NVIDIA-SMI 或者 XPU-SMI 指令，请检查后重试！',
             driverVersion: '驱动版本',
             cudaVersion: 'CUDA 版本',
             process: '进程信息',
diff --git a/frontend/src/views/ai/gpu/index.vue b/frontend/src/views/ai/gpu/index.vue
@@ -30,9 +30,9 @@
                 </div>
             </el-card>
         </div>
-        <el-row :gutter="7" class="card-interval" v-if="options.length !== 0">
-            <el-col :span="24">
-                <el-card style="overflow: inherit">
+        <el-row :gutter="7" v-if="options.length !== 0">
+            <el-col v-bind="gpuType === 'gpu' ? fullWidthProps : halfWidthProps">
+                <el-card class="card-interval" style="overflow: inherit">
                     <template #header>
                         <div :class="mobile ? 'flx-wrap' : 'flex justify-between'">
                             <span class="title">{{ $t('monitor.gpuUtil') }}</span>
@@ -51,7 +51,7 @@
                 </el-card>
             </el-col>
             <el-col :xs="24" :sm="24" :md="12" :lg="12" :xl="12">
-                <el-card style="overflow: inherit">
+                <el-card class="card-interval" style="overflow: inherit">
                     <template #header>
                         <div :class="mobile ? 'flx-wrap' : 'flex justify-between'">
                             <span class="title">{{ $t('monitor.memoryUsage') }}</span>
@@ -70,7 +70,7 @@
                 </el-card>
             </el-col>
             <el-col :xs="24" :sm="24" :md="12" :lg="12" :xl="12">
-                <el-card style="overflow: inherit">
+                <el-card class="card-interval" style="overflow: inherit">
                     <template #header>
                         <div :class="mobile ? 'flx-wrap' : 'flex justify-between'">
                             <span class="title">{{ $t('monitor.powerUsage') }}</span>
@@ -89,7 +89,7 @@
                 </el-card>
             </el-col>
             <el-col :xs="24" :sm="24" :md="12" :lg="12" :xl="12">
-                <el-card style="overflow: inherit">
+                <el-card class="card-interval" style="overflow: inherit">
                     <template #header>
                         <div>
                             {{ $t('monitor.temperature') }}
@@ -110,8 +110,8 @@
                     </div>
                 </el-card>
             </el-col>
-            <el-col :xs="24" :sm="24" :md="12" :lg="12" :xl="12">
-                <el-card style="overflow: inherit">
+            <el-col :xs="24" :sm="24" :md="12" :lg="12" :xl="12" v-if="gpuType === 'gpu'">
+                <el-card class="card-interval" style="overflow: inherit">
                     <template #header>
                         <div :class="mobile ? 'flx-wrap' : 'flex justify-between'">
                             <span class="title">{{ $t('monitor.fanSpeed') }}</span>
@@ -161,8 +161,12 @@ const mobile = computed(() => {
     return globalStore.isMobile();
 });
 
+const fullWidthProps = { span: 24 };
+const halfWidthProps = { xs: 24, sm: 24, md: 12, lg: 12, xl: 12 };
+
 const loading = ref(false);
 const options = ref([]);
+const gpuType = ref('gpu');
 const timeRangeGlobal = ref<[Date, Date]>([new Date(new Date().setHours(0, 0, 0, 0)), new Date()]);
 const chartsOption = ref({
     loadPowerChart: null,
@@ -189,14 +193,19 @@ const search = async () => {
         .then((res) => {
             loading.value = false;
             options.value = res.data.productNames || [];
+            gpuType.value = res.data.gpuType || 'gpu';
             searchInfo.productName = searchInfo.productName || (options.value.length > 0 ? options.value[0] : '');
             let baseDate = res.data.date.length === 0 ? loadEmptyDate(timeRangeGlobal.value) : res.data.date;
             let date = baseDate.map(function (item: any) {
                 return dateFormatWithoutYear(item);
             });
             initCPUCharts(date, res.data.gpuValue);
             initMemoryCharts(date, res.data.memoryValue);
-            initPowerCharts(date, res.data.powerValue);
+            if (gpuType.value === 'gpu') {
+                initPowerCharts(date, res.data.powerValue);
+            } else {
+                initXpuPowerCharts(date, res.data.powerValue);
+            }
             initSpeedCharts(date, res.data.speedValue);
             initTemperatureCharts(date, res.data.temperatureValue);
         })
@@ -270,6 +279,33 @@ function initPowerCharts(baseDate: any, items: any) {
         formatStr: '%',
     };
 }
+
+function initXpuPowerCharts(baseDate: any, items: any) {
+    let list = items.map(function (item: any) {
+        return { value: Number(item.used.toFixed(2)), data: item };
+    });
+    list = list.length === 0 ? loadEmptyData2() : list;
+    chartsOption.value['loadPowerChart'] = {
+        xData: baseDate,
+        yData: [
+            {
+                name: i18n.global.t('monitor.powerUsage'),
+                data: list,
+            },
+        ],
+        tooltip: {
+            trigger: 'axis',
+            formatter: function (list: any) {
+                let res = loadDate(list[0].name);
+                for (const item of list) {
+                    res += loadSeries(item, item.data.value ? item.data.value : item.data, 'W');
+                }
+                return res;
+            },
+        },
+        formatStr: 'W',
+    };
+}
 function initTemperatureCharts(baseDate: any, items: any) {
     let temperatures = items.map(function (item: any) {
         return Number(item);
@@ -334,14 +370,15 @@ function withMemoryProcess(list: any) {
     if (!process) {
         return res;
     }
+    let title = gpuType.value === 'gpu' ? i18n.global.t('aiTools.gpu.type') : i18n.global.t('aiTools.gpu.shr');
     res += `
         <div style="margin-top: 10px; border-bottom: 1px dashed black;"></div>
         <table style="border-collapse: collapse; margin-top: 20px; font-size: 12px;">
         <thead>
             <tr>
             <th style="padding: 6px 8px;">PID</th>
-            <th style="padding: 6px 8px;">${i18n.global.t('aiTools.gpu.type')}</th>
             <th style="padding: 6px 8px;">${i18n.global.t('aiTools.gpu.processName')}</th>
+            <th style="padding: 6px 8px;">${title}</th>
             <th style="padding: 6px 8px;">${i18n.global.t('aiTools.gpu.processMemoryUsage')}</th>
             </tr>
         </thead>
@@ -354,10 +391,10 @@ function withMemoryProcess(list: any) {
                     ${row.pid}
                 </td>
                 <td style="padding: 6px 8px; text-align: center;">
-                    ${loadProcessType(row.type)}
+                    ${row.processName}
                 </td>
                 <td style="padding: 6px 8px; text-align: center;">
-                    ${row.processName}
+                    ${loadProcessType(row.type)}
                 </td>
                 <td style="padding: 6px 8px; text-align: center;">
                     ${row.usedMemory}
diff --git a/frontend/src/views/home/status/index.vue b/frontend/src/views/home/status/index.vue
@@ -69,8 +69,9 @@
                     <el-button v-if="!cpuShowAll" @click="cpuShowAll = true" icon="More" link size="small" />
                     <el-button v-if="cpuShowAll" @click="cpuShowAll = false" icon="ArrowUp" link size="small" />
                 </div>
+                <br />
 
-                <el-button link size="small" type="primary" class="mt-2 mb-2" @click="showTop = !showTop">
+                <el-button link size="small" type="primary" class="mt-1 mb-2" @click="showTop = !showTop">
                     {{ $t('home.cpuTop') }}
                     <el-icon v-if="!showTop"><ArrowRight /></el-icon>
                     <el-icon v-if="showTop"><ArrowDown /></el-icon>
@@ -315,7 +316,7 @@
                 <span class="input-help" v-else>{{ item.deviceName }}</span>
             </el-col>
         </template>
-        <el-col :xs="6" :sm="6" :md="6" :lg="6" :xl="6" align="center" v-if="totalCount > 5">
+        <el-col :xs="6" :sm="6" :md="3" :lg="3" :xl="3" align="center" v-if="totalCount > 5">
             <el-button v-if="!showMore" link type="primary" @click="changeShowMore(true)" class="buttonClass">
                 {{ $t('tabs.more') }}
                 <el-icon><Bottom /></el-icon>

Original file line number	Diff line number	Diff line change
`@@ -721,7 +721,7 @@ var UpdateDatabase = &gormigrate.Migration{`
`721`	`721`	`}`
`722`	`722`
`723`	`723`	`var AddGPUMonitor = &gormigrate.Migration{`
`724`		`- ID: "20251119-add-gpu-monitor",`
	`724`	`+ ID: "20251127-add-gpu-monitor",`
`725`	`725`	`Migrate: func(tx *gorm.DB) error {`
`726`	`726`	`return global.GPUMonitorDB.AutoMigrate(&model.MonitorGPU{})`
`727`	`727`	`},`
Original file line number	Diff line number	Diff line change
`@@ -167,6 +167,7 @@ export namespace Host {`
`167`	`167`	`endTime: Date;`
`168`	`168`	`}`
`169`	`169`	`export interface MonitorGPUData {`
	`170`	`+ gpuType: string;`
`170`	`171`	`productNames: Array<string>;`
`171`	`172`	`date: Array<Date>;`
`172`	`173`	`gpuValue: Array<number>;`